The Journal of Neuroscience, December 24, 2008 28(52):1422314229 14223

Development/Plasticity/Repair

Prolonged Perceptual Learning of Positional Acuity in Adult

Amblyopia: Perceptual Template Retuning Dynamics
Roger W Li,1,2 Stanley A Klein,1 and Dennis M Levi1
1School of Optometry and the Helen Wills Neuroscience Institute, University of California, Berkeley, Berkeley, California 94720-2020, and 2Department of
Psychology, The University of Hong Kong, Pokfulam, Hong Kong

Amblyopia is a developmental abnormality that results in physiological alterations in the visual cortex and impairs form vision. It is often
successfully treated by patching the sound eye in infants and young children, but is generally considered to be untreatable in adults.
However, a number of recent studies suggest that repetitive practice of a visual task using the amblyopic eye results in improved
performance in both children and adults with amblyopia. These perceptual learning studies have used relatively brief periods of practice;
however, clinical studies have shown that the time-constant for successful patching is long. The time-constant for perceptual learning in
amblyopia is still unknown. Here we show that the time-constant for perceptual learning depends on the degree of amblyopia. Severe
amblyopia requires 50 h (35,000 trials) to reach plateau, yielding as much as a five-fold improvement in performance at a rate of
1.5%/h. There is significant transfer of learning from the amblyopic to the dominant eye, suggesting that the learning reflects alter-
ations in higher decision stages of processing. Using a reverse correlation technique, we document, for the first time, a dynamic retuning
of the amblyopic perceptual decision template and a substantial reduction in internal spatial distortion. These results show that the
mature amblyopic brain is surprisingly malleable, and point to more intensive treatment methods for amblyopia.
Key words: plasticity; critical period; visual learning; positional acuity; classification image; amblyopia

Introduction adults with amblyopia and that this improvement transferred to

Amblyopia (from the Greek, amblyos: blunt; opia: vision) is a
developmental disorder that results in physiological alterations in
the visual cortex and impairs form vision (Levi et al., 1991), and is
usually associated with abnormal visual experience caused by
strabismus, anisometropia and visual deprivation early in life.
Aside from refractive error, amblyopia is one of the main causes
for the loss of vision in children (Newman and East, 2000). Clin-
ically, it is often successfully treated by patching the sound eye in
infants and young children (occlusion therapy), but is generally
considered to be untreatable in adults and older children who fall
outside the period of cortical plasticity. A long period of occlu-
sion, ranging from several months to a year, is needed for the
recovery of amblyopic vision, with acuity improving 0.22% per
hour of occlusion (Stewart et al., 2004), and the doseresponse
appears to plateau only after 100 400 h (Cleary, 2000; Stewart et
al., 2004, 2005).
Although treatment of amblyopia is generally not applied to
adults, it is now clear that adults with amblyopia can improve
following perceptual learning, i.e., repetitive practice of a de-
manding visual task (Levi and Li, 2008). Over a decade ago, Levi
and Polat first showed that practice enhanced Vernier acuity in
Received Sept. 7, 2008; accepted Nov. 8, 2008.
This work was supported by National Eye Institute Grants R01EY01728 (D.M.L.) and R01EY04776 (S.A.K.) and a
James S. McDonnell Foundation grant (D.M.L.).
Correspondence should be addressed to Roger Li, School of Optometry, University of California, Berkeley, Berke-
ley, CA 94720-2020. E-mail: oroger@berkeley.edu.
DOI:10.1523/JNEUROSCI.4271-08.2008
Copyright 2008 Society for Neuroscience 0270-6474/08/2814223-07$15.00/0
improvement in visual acuity in some of their observers (Levi and
Polat, 1996; Levi et al., 1997). It is now clear that perceptual
learning improves amblyopic performance on a wide range of
tasks, including position discrimination (Li and Levi, 2004), spa-
tial interaction (Polat et al., 2004), contrast detection (Zhou et al.,
2006), and letter recognition (Levi, 2005; Chung et al., 2006,
2008).
Most previous learning studies used relatively brief periods of
practice (Li and Levi, 2004; Li et al., 2005) as it is generally be-
lieved that the adult brain has limited plasticity and thus percep-
tual learning in normal foveal vision is rapidly completed after a
few hours of practice. For example, Levi and Polat (Levi and
Polat, 1996; Levi et al., 1997) had observers complete only 4000
5000 practice of trials (510 h). Li and Levi asked seven ambly-
opic patients to practice a positional discrimination task and they
seemed to have reached the plateau level and showed a modest
improvement (20 30%) in a brief period (20 h) of training (Li
and Levi, 2004; Li et al., 2005). Other studies used a criterion of
asymptotic performance on three consecutive sessions, giving a
mean length of training of 12 sessions, or 12 h (Zhou et al.,
2006; Huang et al., 2008). The most extensive training was pro-
vided by Polat et al. Their subjects practiced a contrast detection
task for, on average 45 sessions, 22 h (Polat et al., 2004). Inter-
estingly, close inspection of their data, as well as that of the other
studies, suggests that observers may still have not reached their
ultimate asymptotic performance. This raises the question of
whether prolonged perceptual learning would result in more
substantial improvement in adults with amblyopia.
A recent study revealed that in two juvenile amblyopes, pro-
14224 J. Neurosci., December 24, 2008 28(52):1422314229 Li et al. Prolonged Perceptual Learning in Adult Amblyopia

Table 1. Clinical data

Observer Age (years) Gender Cover test Eye Refractive error Snellen VA Stereopsis (arcsec) Type of amblyopia
1
SF 18 M 6 L ExoT R 1.50/0.25 90 20/12.5 Failed (400) Strab
L Plano/1.00 30 20/1252 (20/802)
JS 22 F 7 L EsoT R 1.00 20/16 70 Strab
5 L HyperT L 0.75 20/321 (20/201)
AA 29 F 30 Alt EsoT R 2.00/2.25 180 20/252 (20/202) Failed (400) Strab
10 R HyperT L 3.75/2.00 5 20/202
SC 29 M NMD R 0.50 20/162 Failed (400) Aniso
L 3.25/0.50 155 20/502 (20/402)
ED 39 M NMD R 2.25/1.00 70 20/32 (20/321) 25 Aniso
L 0.25/0.25 110 20/162
VC 22 F 5 R ExoT R 4.75/1.50 10 20/1254 (20/1251) Failed (400) Strab and Aniso
L Plano/0.50 170 20/12.52
AW 23 F 5 R ExoT R 2.75/1.00 160 20/801 (20/501) Failed (400) Strab and Aniso
4 L HyperT L 1.00/0.50 180 20/161
BK 24 M NMD R 6.25/1.2510 20/202 20 Normal
L 7.00/1.25 160 20/202
DW 22 F NMD R Plano/0.50 10 20/161 20 Normal
L Plano 20/162
Seven adults with strabismic (Strab) and/or anisometropic (Aniso) amblyopia participated. Their visual acuity (VA) in the amblyopic eye ranged from 20/25 (mild) to 20/125 (severe), and their isolated letter acuity is listed in parenthesis.
M, Male; F, female; R, right; L, left; ExoT, exotropia; EsoT, esotropia; Hyper T, hypertropia; Alt, alternate; NMD, no movement detected.

longed perceptual learning yielded an 2.4-fold improvement in
performance after 50 h of training (Li et al., 2007). One might
argue that the developing brain is more malleable, thus a longer
period of training can produce more substantial improvements.
But it raises the important question of whether or not similar
prolonged perceptual learning would result in more substantial
gains in adult amblyopes. We know that practicing position dis-
crimination can reduce spatial distortion (internal positional
noise) and enhance the ability to extract stimulus information
(sampling efficiency) during the course of training (Li and Levi,
2004; Li et al., 2005, 2007). However, there remains a long-
standing fundamental question of how the amblyopic brain re-
tunes its abnormal neuronal connections so as to optimize and
sharpen its responses to the visual world.
To address these questions, we adopted a much stricter stop-
ping rule than in previous studies to decide whether training
should continue, and our experiment was designed to explore the
amount of experience-dependent plasticity in the mature ambly-
opic brain. Position discrimination depends strongly on spatial
relations, and is thought to be limited by positional uncertainty in
amblyopes. Thus, in this study we used positional noise, i.e.,
perturbation of the positions of parts of the stimulus, to mimic
the putative limiting internal noise by rendering the positions of
the samples uncertain (rather than by obscuring their visibility, as
occurs with luminance noise) and to explore the underlying neu-
ral mechanisms for position discrimination. Our results reveal
surprisingly substantial cortical plasticity in adult amblyopia fol-
lowing a prolonged period of repetitive visual stimulation. Using
a reverse correlation technique, we were able to document, for
the first time, the session-to-session retuning dynamics of the
perceptual decision template or perceptive fields of the ambly-
opic brain during visual learning. An understanding of the limits,
time course, and mechanisms of perceptual learning is critical for
developing a more effective amblyopia treatment.
Materials and Methods
Experimental procedures. Seven adults aged 18 40 years with anisome-
tropic and/or strabismic amblyopia participated in the study (Table 1).
They practiced the positional discrimination task repeatedly using their
amblyopic eye; the other eye was occluded with a standard black eye
patch. Each days session consisted of 960 trials in 1.5 h, 35 sessions a
week.
To evaluate the possible maximum limits of cortical plasticity, we
adopted a much stricter criterion for ending the training course than
those in our previous study (Li and Levi, 2004); our observers were
required to continue practicing the visual task until their improved per-
formance were stable and maintained for as much as 1520 sessions after
the plateau levels were reached. Consequently, three of our observers had
given 50,000 responses in 50 training sessions. All observers were naive
to the purpose of experiment.
Visual stimulus. The stimulus was comprised of two horizontal seg-
ments with a 1.25 gap between the centers of the innermost patches (see
Fig. 1 A). Each segment consisted of five Gabor patches, and the inter-
patch separation was 0.83. The patches were constructed to have a 2/3
aspect ratio: the Gaussian envelope SD was 0.39 and 0.58 for the hori-
zontal and vertical orientations respectively. Positional noise was pro-
duced by shifting the position of each Gabor patch in the vertical direc-
tion around the intended mean line position of the test (right) segment
according to a discrete binary probability function. The binary noise
amplitude was always 0.11, either positive (downward) or negative (up-
ward). No noise was added to the reference (left) segment. The carrier
spatial frequency, and stimulus size, was scaled [5 (VC and AW), 6.6 (SF),
or 10 (SC, JS, ED, AA, and normal observers) cycles per degree] accord-
ing to observers baseline visual acuity of the amblyopic eye by varying
viewing distance (4, 2, or 1m), without changing any physical dimensions
of the stimulus displayed on the monitor screen.
The stimulus was briefly presented (200 ms) on a flat 21-inch Sony
F520 monitor screen at 90 Hz refresh rate. Subjects were asked to main-
tain fixation at the center of the monitor screen. The mean luminance of
the stimuli was 55 cd/m 2, and the contrast of each Gabor patch was 84%.
Threshold calculation. On each trial, the test line was presented ran-
domly in one of three positions: aligned with the reference line, or one
step above () or below () it. The observers task was to rate the
position of the test segment compared with the reference segment by
giving an integer number from 2 (above) to 2 (below), including 0
(aligned). Trial-by-trial verbal feedback was provided following each
trial. Observers were instructed to attend to all five Gabors in each seg-
ment so as to determine the average segment positions and detect the
direction of misalignment. A rating-scale signal detection paradigm was
used to calculate d (detectability or effect size) for discriminating the
direction of offset (Levi et al., 2000). The position offset at which d 1
was taken as threshold. At the start of each training session, the offset
between the two segments was adjusted according to the previous ses-
Li et al. Prolonged Perceptual Learning in Adult Amblyopia J. Neurosci., December 24, 2008 28(52):1422314229 14225

Figure 1. Prolonged perceptual learning of position discrimination in adult amblyopia. A, Visual stimulus. The observers task was to identify the location of the right test segment relative to the
left reference segment. Positional noise was added to the right segment by jittering the vertical position of each of the five Gabor patches (either up or down). B, Learning profile. Each data point
represents one session consisting of 960 trials in 1.5 h. It is worth noting that it took as many as 35 sessions (35,000 trials) to improve and reach a stable performance plateau (prolonged
learning). Observer VC withdrew from the study after 19 sessions (open symbols in this and subsequent figures) and did not finish the experiment according to our end-point criterion. Gray symbols
indicate the mean data of two normal observers for comparison. C, The number of hours needed to reach stable plateau performance as a function of pretraining positional threshold. D, Acuity
improvement (pre/post ratio) as a function of pretraining positional threshold. The data of two children (black symbols in panels C and D) with amblyopia who underwent intensive perceptual
learning are replotted here (Li et al., 2007). E, Comparison of positional acuity between the two eyes, nonamblyopic (NAE) and amblyopic (AE). The posttraining performance (filled symbols) is much
closer to (or even better than) the gray 1:1 reference line. The amount of plateau improvement, as indicated by the length of colored lines, is considerably larger for those observers with much
elevated baseline pretraining threshold (open symbols). F, Interocular transfer (from the trained AE to the nontrained fellow NAE) versus phase II direct training (on the previously untrained NAE).
G, Improvement in visual acuity and positional acuity. The amount of improvement in visual acuity is not dependent on that in positional acuity (slope 0.0015 0.04, t 0.04, p 0.97).

sions thresholds so as to keep the detectability around 1, aiming to make
the task challenging, not too easy or too difficult.
The decision template and internal noise. The details of decision tem-
plate (also known as classification image) and internal noise calculations
are fully described elsewhere (Li et al., 2006). This template provides the
decision rules of how the brain extracts and integrates stimulus informa-
tion for positional judgments. In brief, a multiple linear regression was
used to reverse correlate the individual positions of the 5 Gabors (see Fig.
1 A) with the human responses (2 to 2) in the computation of per-
ceptive fields, with the sum of the five normalized weightings equal to
unity. For each of three offset levels (, 0, and ), there are 2 5 32
possible external noise combinations in total.
Each stimulus arrangement was presented 10 times to observers,
therefore each session consisted of 960 trials (32 noise combinations 3
offset levels 10 passes) in a random sequence. The 10-pass response
consistency for each noise combination provides a measure of internal
noise. Importantly, the use of binary noise allows us to monitor fewer
moving parts, thus requiring much fewer trials when combined with
multiple regression to compute perceptive fields.
Transfer experiment. At the commencement of the training experi-
ment, we also tested the baseline performance of the fellow nonambly-
opic eye at the same viewing distance as selected for testing the amblyopic
eye. After Phase 1 training on the amblyopic eye, we switched to train the
nonamblyopic eye in Phase 2 training to evaluate whether there was any
interocular transfer (complete or incomplete) from the trained ambly-
opic eye to the untrained fellow eye.
Clinical vision tests. In this study, a Bailey-Lovie logMAR letter chart
(National Vision Research Institute of Australia, 1978) was used to mea-
sure visual acuity, and Randot Stereotest (Stereo Optical Company) was
used to measure stereoacuity for the range of 20 400 arcsec.
Results
Prolonged perceptual learning in adult amblyopia
Most previous studies (Levi and Polat, 1996; Li and Levi, 2004;
Dosher and Lu, 2007) have used either a fixed duration of prac-
tice (typically 10 20 h or 4000 8000 trials) or a stopping crite-
rion of asymptotic performance on three consecutive sessions
(Zhou et al., 2006). However, our results suggest the need for a
much stricter stopping rule. Extended perceptual learning results
in a more substantial, although not complete recovery in posi-
tional threshold (Fig. 1 B). Practice improves positional acuity
14226 J. Neurosci., December 24, 2008 28(52):1422314229
(the minimum offset between the two segments to be detected
decreases across training sessions) gradually until plateau perfor-
mance is obtained. A three parameter two-line segment curve was
used to fit the logarithmic threshold data as a function of training
session; the intersection point of the two lines indicates the pla-
teau performance for each observer. Note that the threshold data
are plotted in Gabor wavelength () units to facilitate the com-
parison between the different viewing distances used to test indi-
vidual observers. Within this extended period of improvement
there may be short plateaus in performance (Fig. 1 B, inset), fol-
lowed by further improvement.
Interestingly, observers with mild amblyopia improve at a
faster rate than those with severe amblyopia, as reflected by a
steeper regression slopes in Figure 1 B [mean improvement rate:
mild (JS and AA) 11.4 1.9%/h; severe (SF, VC, SC, and AW)
4.0 0.7%/h]. More practice is needed to reach asymptote in
severe amblyopia than in mild amblyopia (Fig. 1C), and the
amount of improvement is directly proportional to the severity of
amblyopia on a log-log scale (Fig. 1 D). The pre/post acuity ratio
was larger than a factor of 4 in a severe amblyope (SF). In general,
deep amblyopes [SF, SC and AW (visual acuity, VA): 20/50 20/
125)] required 50 h (35,000 trials) to reach plateau, and the
acuity ratio was as much as 1.75 4.23-fold. In contrast, mild
amblyopes [ED, JS and AS (VA: 20/2520/40)] required fewer
practice trials (7.530 h) to obtain stable improvement (1.35
1.75-fold). The exponential time-constants were 18.75 and 5.55 h
for deep and mild amblyopes respectively and just 2.7 h for nor-
mal control observers. The black symbols in Figures 1, C and D,
show data of two children (ages 8 and 11) with amblyopia who
underwent intensive perceptual learning. While the stimulus
configurations and psychophysical methods were different, these
data provide a close match with the adult data, suggesting that age
(at least from 8 to 39) is not an important determinant of the
outcome of perceptual learning.
The posttraining threshold in the amblyopic eye (solid sym-
bols) approaches, and even crosses the 1:1 line representing
equality of threshold in the amblyopic and nonamblyopic eyes
(Fig. 1 E). Note that one observer (VC, dashed line) did not com-
plete the training according to our stopping criterion; thus it is
possible that additional practice could result in further improve-
ment. Interestingly in three mild amblyopes (AA, JS, and ED),
plateau performance of the trained amblyopic eye was even better
than the baseline performance of the untrained normal eye. In
other words, the amblyopic deficits for this type of positional
acuity are more or less eliminated.
An important question is whether the improved performance
is maintained after cessation of training. We retested four observ-
ers on the same task after a period of 12 months and found that
all or most of the improved performance was retained (the right-
most symbols of Fig. 1 B). Although there was a small elevation in
threshold (6.5% 1.8% when compared with the baseline mea-
surement), only a few (5) maintenance sessions were enough to
achieve the original plateau levels. Note that after direct training
on the amblyopic eye, two of those four observers (SC and JS)
elected to proceed in training the nonamblyopic eye for a month
(interocular transfer experiment) immediately before retesting
the amblyopic eye. One might speculate that the learning effects,
if any, in the nonamblyopic eye could possibly transfer, to a cer-
tain extent, back to the amblyopic eye. Therefore, we retested the
maintenance performance in another two observers (SF and ED)
who did not participate in the transfer experiment 12 months
after the last session of visual testing and confirmed that the
Li et al. Prolonged Perceptual Learning in Adult Amblyopia
improved performance was long-lasting and substantially
maintained.
The roles of higher- and lower-level processing in learning
One way to reveal where neural alternations occur is to evaluate
interocular transfer of learning. An absence of transfer from the
trained eye to the untrained eye is often taken as evidence for
neural alternations occurring at a stage of visual processing be-
fore binocular convergence. Alternatively, a complete transfer
would be expected if higher cortical areas are involved. Our find-
ings (Fig. 1 F) show a significant transfer of learning to the sound
eye in 3 of the four observers (mean pre/post acuity ratio
1.27 0.11, t 2.458, p 0.046). Interestingly, with further
practice the nonamblyopic eye showed additional improvement,
mean pre/post acuity ratio 1.68 0.16 (mean difference in
acuity ratio 0.41 0.08, paired t 4.995, p 0.015). This
suggests that training the amblyopic eye modified the response
characteristics of higher visual cortex which in turn results in the
interocular transfer, and the subsequent training of the fellow eye
itself may have further strengthened early neural mechanisms.
Recent physiological studies also reported the involvement of
higher decision stage in visual learning (Law and Gold, 2008).
These effects may also explain the accompanying generalized
transfer to improved visual (letter) acuity (Fig. 1G), the sine qua
non of amblyopia, consistent with the recent finding that percep-
tual learning has a broad bandwidth in amblyopia (Huang et al.,
2008).
Retuning dynamics of behavioral receptive fields
Using an efficient trial-by-trial reverse correlation method (Li et
al., 2004, 2006), we are able to monitor changes, if any, in the
decision template or behavioral receptive field (Gold et al.,
1999) for positional judgments during visual learning. The deci-
sion template reveals how observers weight and integrate the po-
sitional information of the 5 patches comprising the test segment
of the visual stimulus (Fig. 1 A). An ideal observer who knows
the stimulus exactly would give equal weight (0.2) to each patch
for positional averaging (Fig. 2 A, inset). Not surprisingly, human
observers adopt a less efficient template for extracting positional
information; the averaging computation relies on a limited num-
ber of stimulus samples.
Importantly, initially the amblyopic eye shows distinct differ-
ences in positional integration from the fellow sound eye (Fig.
2 B): (1) The peak of tuning is shifted to the right (away from the
inner segment) in observers SC (patch 4), SF (patch 5) and VC
(patch 5). In fact, SFs and VCs responses are mostly based on the
position of the outermost patch. This may reflect abnormal sup-
pression, or crowding (Nandy and Tjan, 2007; Levi, 2008) effects
of the reference (left) segment on the proximal samples of the test
(right) segment. In contrast, their nonamblyopic eyes used a
more efficient ideal-observer-like template. (2) With our abut-
ting stimulus, one might expect that amblyopic observers would
down-weight the first (inner) patch because of crowding effects.
In contrast, we observed a very strong response, but in the wrong
direction: repulsion error [patch 1: SC (0.37), SF (1.11), and
VC (0.83); ideal observer (0.2)], representing a misperception
of position (or spatial distortion): when the patch is up, the per-
ceived position is down, and vice versa. It is widely known that
spatial interaction affects contrast (Polat et al., 2005) and contour
(Hess et al., 2001) perception, and here is another example of
abnormal interaction for positional perception. (3) In an extreme
case, severe amblyopes SF and VC adopted an ideal-observer like
template [note that the gray open (VC, 80.6% efficiency as illus-
Li et al. Prolonged Perceptual Learning in Adult Amblyopia J. Neurosci., December 24, 2008 28(52):1422314229 14227

Figure 2. Neural mechanisms involved in learning position discrimination. A, Inset, Ideal versus human template. In human observers (n 2), the averaging computation mostly relied on a few
stimulus samples; much higher weights were given to the selected samples than to the others (red line). Following practice, the retuned template (blue line) is closer to resembling the ideal template
(black line). A, The enhancement of template efficiency after perceptual learning. In four deeply amblyopic observers, the posttraining efficiency is much higher than the pretraining efficiency.
Nevertheless, the other three mild amblyopes (ED, JS, and AA) did not show any significant retuning, as shown by similar pretraining and posttraining efficiency along a 1:1 reference line. B,
Template retuning. A red line illustrates the pretraining template and a blue line illustrates the posttraining template instead. The retuned template is more similar to the ideal template in shape
(black line). For comparison, the data of the nonamblyopic eye (NAE) is plotted as a gray line. C, Perceptual learning results in a decrease in random internal (int.) noise. It appears that the amount
of decrement is directly proportional to the baseline noise level. Note that observer VC (open symbols) did not complete the experiment according to our plateau criterion.

trated in Fig. 2 A) and solid (SF, 67.4% efficiency) symbols fall
close to the black horizontal line for ideal perceptual fields] in the
fellow sound eyes, but the least efficient template of only 6%
efficiency (a very tilted oblique template profile placing unequal
weights for different parts of stimulus with an abnormal very
large repulsion error at one end) in the amblyopic eyes. (4) How-
ever, two other amblyopes with mild loss in visual acuity (JS and
AA) showed nearly normal templates (efficiency: 70 80%)
which are very comparable with the two normal observers (effi-
ciency: 75%).
Most importantly, we recorded a dramatic optimization of the
decision template which, in part, explains the enhanced perfor-
mance following practice, making the posttraining template (Fig.
2 B) more similar in shape to both the nonamblyopic eye and the
ideal template. On average, the template efficiency (squared cor-
relation between human and ideal templates) improved substan-
tially by 126.3% for those four deep amblyopes (Fig. 2 A). The
retuned template is capable of more effective sampling with
much reduced repulsion [e.g., SF: 1.1230.42 (patch 1)], re-
flecting how the brain recalibrates the weightings of neuronal
connections with response feedback to use the spatial informa-
tion from lower-level visual mechanisms more effectively and
appropriately. In contrast, three mild amblyopes did not show
any enhancement in template efficiency because they were al-
ready highly efficient (mean: 72.5%, quite similar to normal con-
trols 74.9%), leaving little room for improvement.
Visual learning also results in a substantial reduction in inter-
nal positional noise (random jitter). We assayed internal noise by
repeating each combination of signal and noise 10 times in each
session. Since the signals and noise are identical, any inconsis-
tency in responses provides an accurate assay of internal noise
(Green, 1964). All observers showed a reduction of internal noise
(mean: 40.2%) in the posttraining session (expressed in unit in
Fig. 2C to facilitate the comparison between different viewing
distances used for individual observers), ranging from 18.4%
(AA) to 55.2% (SF). The more severe the amblyopia, the larger
the reduction (farther away from the 1:1 reference line).
We can model the observers performance, human threshold
th (Fig. 3B, black line), by summing the various sources of noise
(th
2
temp
2
random
2
systemic
2
): template noise (temp) due to
a nonideal decision template and random (random) and system-
atic (systematic) internal noise beyond the template (Li et al.,
2006). The data are replotted in Figure 3A to summarize the
changes in template efficiency and random internal noise. Note
that random internal noise is specified as pre/post ratio instead of
post/pre ratio for template efficiency so that a ratio of 1 repre-
sents increased template efficiency or decreased internal noise.
Figure 3B, which shows the squared threshold of each model
component, reveals the dynamics of template retuning and inter-
nal noise reduction across sessions for two observers. The thresh-
old of the ideal template, ideal, is 0.298 arcmin (gray area: 0.148
at 4m viewing distance) reflecting averaging error because of
stimulus positional uncertainty; the human template efficiency is
defined as E f fTemp (ideal/temp) 2 100%. It is especially
interesting to see the very gradual template retuning in observer
SC (yellow area: template threshold gradually decreased until 40 45
h of practice; template efficiency improved substantially from 30%
to 70%), resulting in a very efficient template, only slightly less effi-
cient than the ideal template after 35-thousand trials. In contrast, the
improvement in observer ED can be mainly attributed to the re-
duced internal random noise (pre/post ratio: 1.58), the yellow area
remains flat (pre/post ratio: 0.97) across the training course. The
right end of each panel shows the maintenance data; it is clear that all
threshold components remain much the same.
Discussion
After a prolonged period of perceptual learning positional acuity
in adult amblyopia can be substantially normalized, with as much
14228 J. Neurosci., December 24, 2008 28(52):1422314229
Figure 3. Retuning dynamics of behavioral receptive fields. A, Summary of changes in template efficiency and internal random noise. Template efficiency is expressed as post/pre ratio, while
internal noise is expressed as pre/post ratio. B, Threshold components. The squared sum of all threshold components (ideal template, human template, random noise, and systematic noise) is quite
close to the squared human threshold, although our modeling tends to slightly overestimate human performance, as indicated by a black line. In agreement with our earlier study with normal
observers (Li et al., 2006), systematic noise (like the systematic template error, but related to a higher-order nonlinearity) is negligible in this positional task.
as a five-fold improvement in performance in observers with
deep amblyopia. A much longer time than expected is necessary
to reach a real plateau level (50 h). In normal adults, in agree-
ment with our earlier study (Li et al., 2004), visual learning is much
faster, mostly completed in just a few sessions totaling 510 h (Fig.
1B, gray symbols). Our findings clearly show that the mature ambly-
opic brain, which is generally thought to have limited capacity for
neural alternations, shows substantial plasticity with prolonged per-
ceptual learning. To our knowledge, such prolonged learning result-
ing in such substantial gains in visual performance, have not been
previously reported in adults with normal or amblyopic vision. Im-
portantly, the learning effects are long-lasting (Fig. 1B), and only a
few retraining sessions are needed to compensate for a slight loss of
previously gained improvement.
One might argue that the improvement in performance might be
in part attributed to high level cognitive task learning or instrument
learning (Westheimer, 2001). For example, amblyopes might learn
to fixate and/or accommodate more accurately with their amblyopic
eye. We argue that improvements in fixation and/or focusing accu-
racy are unlikely to account for our present findings for several rea-
sons: (1) Our task (long and very visible horizontal pair of Gabor
patch groupings with low/moderate carrier spatial frequency) is not
strongly dependent on precise fixation or focus. (2) The very grad-
ual, but substantial, improvement in thresholds across an extended
period of training sessions indicates that the changes in sensitivity are
genuine. (3) If the improvement in performance were due to im-
proved fixation/focus, we would expect the improvement to transfer
across orientations. However our previous studies using a similar
positional task found no transfer across orthogonal orientations (Li
and Levi, 2004).
A stringent stopping rule is important for maximizing the
treatment effects. As an example, when only the first 10-sessions
of SFs data are fitted with an exponential function (Fig. 1 B,
inset), it appears that the plateau performance (improvement:
51%) would have been obtained at around 7 8 h. But further
practice resulted in an additional 52% improvement over 40
more hours. In contrast to cognitive task learning (Westheimer,
2001) that is rapid, this gradual improvement reflects genuine
experience-dependent neural alternations. Similar prolonged
learning effects have also been recently reported in animals with
degraded auditory systems (Zhou and Merzenich, 2007).
Li et al. Prolonged Perceptual Learning in Adult Amblyopia
An earlier study failed to report any correlation between the
depth of amblyopia and the amount of improvement (Levi et al.,
1997), most likely due to incomplete learning after just 10 prac-
tice sessions. Those previous studies (Levi and Polat, 1996; Levi et
al., 1997; Li and Levi, 2004) might have largely under-estimated
the limits of neural plasticity in the mature amblyopic brain.
Importantly, we quantified the changes in the observers percep-
tual decision templates for positional acuity as their position acuity
improved. This perceptual decision template represents a spatial
map that reveals how the lower-end spatial filters are calibrated,
weighted and integrated for position localization. In some sense, the
abnormal template before learning reflects the nature of spatial dis-
tortion and visual crowding in amblyopic vision. Here we provide
evidence that the perceptual decision template of the amblyopic vi-
sual system can be retuned through repetitive practice. Thus, neuro-
nal connections in the amblyopic brain are not completely hard-
wired, but are retunable. The retuned template is less distorted and
more effective in interpreting positional information. Unlike the tra-
ditional averaging methods for calculating perceptive fields that re-
quire several thousands of responses, our efficient reverse correla-
tion technique (Li et al., 2006) allows us to obtain a reliable template
from only one thousand trials and to document the session-by-
session retuning dynamics of the perceptual decision template dur-
ing perceptual learning.
Neuronal connections of the visual system are generally
thought to be not malleable into adulthood, and the question of
reorganization in adult cortex following a retinal scotoma re-
mains controversial (Smirnakis et al., 2005). However, a recent
calcium imaging study has shown that rewiring can indeed occur
in adult cortex following stroke (Winship and Murphy, 2008),
supporting the notion that somatosensory cortical neurons still
retain a remarkable degree of adaptive plasticity. Our results
show that the improvement following practice with the ambly-
opic eye transferred to the sound eye. This result is consistent
with some learning taking place at a higher decision stage of
visual processing. It is worth noting that binocular transfer may
occur as early as V1 layers 23 (Zhang et al., 2005); however a
more central explanation is that at a high level decision stage,
the visual system may learn to rely on (or attend to) a subset of
neurons, that provide more salient signals, without necessitating
physiological modifications at an earlier stage (Mollon and Da-
Li et al. Prolonged Perceptual Learning in Adult Amblyopia
nilova, 1996). This may also explain the fact that there is substan-
tial further learning in the nonamblyopic eye. In the case of am-
blyopia, the cortical image is degraded and distorted in the
amblyopic eye (Sireteanu et al., 2008). Although the stimulus
configurations are the same, the visual brain may need to relearn
a different, much sharper and more finely calibrated cortical im-
age in the sound eye so as to extract the most salient visual fea-
tures or signals for decision making that could, in part, explain
the secondary learning of the same visual task.
Normal peripheral vision has been shown to be similar to
amblyopic foveal vision in several respects and is, therefore, often
used as a model to understand amblyopia (Levi et al., 1984, 1987;
Wilson, 1991). In normal periphery, a wide range of visual func-
tions are degraded relative to the fovea. We postulate that similar
prolonged learning might also exist in normal periphery. Some
evidence for prolonged learning of orientation discrimination in
peripheral vision can be seen in the data of Furmanski and Engel
(2000). This is particularly important for those who have reduced
central vision and must rely on the remaining para-foveal vision.
In summary, our results show that the adult amblyopic brain
retains a surprising degree of neural plasticity that is revealed by
prolonged perceptual learning, and provide for the first time, docu-
mentation of the session-by-session dynamics of template retuning
and noise reduction in amblyopic vision. We believe this prolonged
learning does not occur in normal foveal vision in which visual
learning is quickly completed in only a few hours, and that it has
important potential clinical applications in visual rehabilitation of
children and adults with amblyopia. In particular, our results point
to the need for customizing the length or intensity of treatment
based on the severity of amblyopia. We understand that there are
individual variations in treatment efficacy and hence a large-scale
multicenter clinical trial is necessary to quantify the doseresponse
relationship for different ages of onset, types and depths of amblyo-
pia and optimize the treatment dose and schedule. Our present find-
ings, together with other recent clinical studies, suggest that it may be
the time to reconsider our notions about the critical period of neural
plasticity in amblyopia.
References
Chung ST, Li RW, Levi DM (2006) Identification of contrast-defined letters
in adults with amblyopia benefits from perceptual learning. Vision Res
46:38533861.
Chung STL, Li RW, Levi DM (2008) Learning to identify near-threshold
luminance-defined and contrast-defined letters in observers with ambly-
opia. Vision Res 48:2739 2750.
Cleary M (2000) Efficacy of occlusion for strabismic amblyopia: can an op-
timal duration be identified? Br J Ophthalmol 84:572578.
Dosher BA, Lu ZL (2007) The functional form of performance improvements
in perceptual learning: learning rates and transfer. Psychol Sci 18:531539.
Furmanski CS, Engel SA (2000) Perceptual learning in object recognition:
object specificity and size invariance. Vision Res 40:473 484.
Gold J, Bennett PJ, Sekuler AB (1999) Signal but not noise changes with
perceptual learning. Nature 402:176 178.
Green DM (1964) Consistency of auditory detection judgments. Psychol
Rev 71:392 407.
Hess RF, Dakin SC, Tewfik M, Brown B (2001) Contour interaction in am-
blyopia: scale selection. Vision Res 41:22852296.
Huang CB, Zhou Y, Lu ZL (2008) Broad bandwidth of perceptual learning
in the visual system of adults with anisometropic amblyopia. Proc Natl
Acad Sci U S A 105:4068 4073.
Law CT, Gold JI (2008) Neural correlates of perceptual learning in a senso-
rymotor, but not a sensory, cortical area. Nat Neurosci 11:505513.
Levi DM (2005) Perceptual learning in adults with amblyopia: a reevalua-
tion of critical periods in human vision. Dev Psychobiol 46:222232.
Levi DM (2008) Crowdingan essential bottleneck for object recognition: A
mini-review. Vision Res 48:635 654.
J. Neurosci., December 24, 2008 28(52):1422314229 14229
Levi DM, Li RW (2008) Improving the performance of the amblyopic visual
system. Philos Trans R Soc Lond B Biol Sci. Advance online publication.
Retrieved December 9, 2008. doi:10.1098/rstb.2008.0203
Levi DM, Polat U (1996) Neural plasticity in adults with amblyopia. Proc
Natl Acad Sci U S A 93:6830 6834.
Levi DM, Klein SA, Aitsebaomo P (1984) Detection and discrimination of
the direction of motion in central and peripheral vision of normal and
amblyopic observers. Vision Res 24:789 800.
Levi DM, Klein SA, Yap YL (1987) Positional uncertainty in peripheral and
amblyopic vision. Vision Res 27:581597.
Levi DM, Ciuffreda KJ, Selenow A (1991) Amblyopia: basic and clinical
aspects. Boston: Butterworth-Heinemann.
Levi DM, Polat U, Hu YS (1997) Improvement in vernier acuity in adults
with amblyopia. Practice makes better. Invest Ophthalmol Vis Sci
38:14931510.
Levi DM, Klein SA, Carney T (2000) Unmasking the mechanisms for Ver-
nier acuity: evidence for a template model for Vernier acuity. Vision Res
40:951972.
Li RW, Levi DM (2004) Characterizing the mechanisms of improvement for
position discrimination in adult amblyopia. J Vis 4:476 487.
Li RW, Levi DM, Klein SA (2004) Perceptual learning improves efficiency by re-
tuning the template for position discrimination. Nat Neurosci 7:178183.
Li RW, Young KG, Hoenig P, Levi DM (2005) Perceptual learning improves
visual perception in juvenile amblyopia. Invest Ophthalmol Vis Sci
46:31613168.
Li RW, Klein SA, Levi DM (2006) The receptive field and internal noise for
position acuity change with feature separation. J Vis 6:311321.
Li RW, Provost A, Levi DM (2007) Extended perceptual learning results in
substantial recovery of both positional acuity and visual acuity in juvenile
amblyopia. Invest Ophthalmol Vis Sci 48:5046 5051.
Mollon JD, Danilova MV (1996) Three remarks on perceptual learning.
Spat Vis 10:5158.
Nandy AS, Tjan BS (2007) The nature of letter crowding as revealed by first-
and second-order classification images. J Vis 7:126.
Newman DK, East MM (2000) Prevalence of amblyopia among defaulters
of preschool vision screening. Ophthalmic Epidemiol 7:6771.
Polat U, Ma-Naim T, Belkin M, Sagi D (2004) Improving vision in adult
amblyopia by perceptual learning. Proc Natl Acad Sci U S A
101:6692 6697.
Polat U, Bonneh Y, Ma-Naim T, Belkin M, Sagi D (2005) Spatial interac-
tions in amblyopia: effects of stimulus parameters and amblyopia type.
Vision Res 45:14711479.
Sireteanu R, Baumer CC, Iftime A (2008) Temporal instability in amblyopic
vision: relationship to a displacement map of visual space. Invest Oph-
thalmol Vis Sci 49:3940 3954.
Smirnakis SM, Brewer AA, Schmid MC, Tolias AS, Schuz A, Augath M, In-
hoffen W, Wandell BA, Logothetis NK (2005) Lack of long-term cortical
reorganization after macaque retinal lesions. Nature 435:300 307.
Stewart CE, Moseley MJ, Stephens DA, Fielder AR (2004) Treatment dose-
response in amblyopia therapy: the Monitored Occlusion Treatment of
Amblyopia Study (MOTAS). Invest Ophthalmol Vis Sci 45:3048 3054.
Stewart CE, Fielder AR, Stephens DA, Moseley MJ (2005) Treatment of uni-
lateral amblyopia: factors influencing visual outcome. Invest Ophthalmol
Vis Sci 46:31523160.
Westheimer G (2001) Is peripheral visual acuity susceptible to perceptual
learning in the adult? Vision Res 41:4752.
Wilson HR (1991) Model of peripheral and amblyopic hyperacuity. Vision
Res 31:967982.
Winship IR, Murphy TH (2008) In vivo calcium imaging reveals functional
rewiring of single somatosensory neurons after stroke. J Neurosci
28:6592 6606.
Zhang B, Bi H, Sakai E, Maruko I, Zheng J, Smith EL 3rd, Chino YM (2005)
Rapid plasticity of binocular connections in developing monkey visual
cortex (V1). Proc Natl Acad Sci U S A 102:9026 9031.
Zhou XM, Merzenich MM (2007) Intensive training in adults refines A1
representations degraded in an early postnatal critical period. Proc Natl
Acad Sci U S A 104:1593515940.
Zhou Y, Huang C, Xu P, Tao L, Qiu Z, Li X, Lu ZL (2006) Perceptual
learning improves contrast sensitivity and visual acuity in adults with
anisometropic amblyopia. Vision Res 46:739 750.